finding
active
finding:gemini-web-search-reward-for-first-300-steps-would-cost-344-064-versus-3-818-for-the-probe-pipeline-2-orders-of-magnitudeGemini web-search reward for first 300 steps would cost ~$344,064 versus ~$3,818 for the probe pipeline (~2 orders of magnitude)
Cost computation.
Source paper
extracted_from(2026) · Aaditya Vikram Prasad · Connor Watts · Jack Merullo · Dhruvil Gala +3
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Calibration enables repurposing.
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- Reward pipeline is ~90× cheaper per rewarded intervention than ground-truth Gemini supervisionfinding0.802Cost efficiency.
- Probe-based method is approximately 10× cheaper than gradient-based alternatives ($30 vs $320 once trained)finding0.764Cost efficiency finding: the probe-based approach costs ~$30 vs ~$320 for gradient-based methods after training.
- A trace shows G3-F turning a nominally wasteful overpay into a net score gain due to the multiplicative formula.
- G3-F escalates bidding by nearly 10× from early to late game, adapting to phase.
- Even stylistic distribution.
- Second-highest lift; Gemini Pro is the highest-gated model in the study
- G2.5-FL raises its own bid in over three-quarters of auction rounds.
- Significant downlift for Gemini.